Genetic Artificial Intelligence in Gastrointestinal Disease

Abstract

The application of predictive and explainable artificial intelligence to bioinformatics data such as single nucleotide polymorphism (SNP) information is attracting rising attention in the diagnosis of various diseases. However, there are few reviews available on the recent progress of genetic artificial intelligence for the early diagnosis of gastrointestinal disease (GID). The purpose of this study is to complete a systematic review on the recent progress of genetic artificial intelligence in GID. The source of data was ten original studies from PubMed. The ten original studies were eligible according to the following criteria: (participants) the dependent variable of GID or associated disease; (interventions/comparisons) artificial intelligence; (outcomes) accuracy, the area under the curve (AUC), and/or variable importance; a publication year of 2010 or later; and the publication language of English. The performance outcomes reported varied within 79–100 for accuracy (%) and 63–98 for the AUC (%). Random forest was the best approach (AUC 98%) for the classification of inflammatory bowel disease with 13 single nucleotide polymorphisms (SNPs). Similarly, random forest was the best method (R-square 99%) for the regression of the gut microbiome SNP saturation number. The following SNPs were discovered to be major variables for the prediction of GID or associated disease: rs2295778, rs13337626, rs2296188, rs2114039 (esophageal adenocarcinoma); rs28785174, rs60532570, rs13056955, rs7660164 (Crohn’s disease early intestinal resection); rs4945943 (Crohn’s disease); rs316115020, rs316420452 (calcium metabolism); rs738409_G, rs2642438_A, rs58542926_T, rs72613567_TA (steatotic liver disease); rs148710154, rs75146099 (esophageal squamous cell carcinoma). The following demographic and health-related variables were found to be important predictors of GID or associated disease besides SNPs: age, body mass index, disease behavior, immune cell type, intestinal microbiome, MARCKS protein, smoking, and SNP density/number. No deep learning study was found even though deep learning was used as a search term together with machine learning. Genetic artificial intelligence is effective and non-invasive as a decision support system for GID.

1. Introduction

1.1. Gastrointestinal Disease

Gastrointestinal disease (GID) is a significant contributor to the global disease burden [1,2,3,4,5,6]. It encompasses the gastrointestinal tract, such as the esophagus, liver, stomach, small and large intestines, gallbladder, and pancreas [1]. GID is responsible for 8 million deaths worldwide each year [2] and 120 billion USD of total expenditure in the United States as of 2018 [3]. In Korea, GID ranked 8th among 21 disease groups in terms of disability-adjusted life years, with 1730 per 100,000 (5.9%) in 2015 [4]. The medical costs of GID in Korea amounted to 4 billion USD or 13% of all medical costs in the country in 2007 [5]. Various factors can contribute to the development of GID, including poor health behaviors like excessive stress, disrupted routine, insufficient exercise, a low-fiber diet, and a high-dairy diet. Unhealthy bowel habits, excessive anti-diarrheal/antacid medication, and pregnancy can also be causes of GID [6]. There are two types of GID: functional and structural. Functional GID is characterized by the normal appearance of the gastrointestinal tract but motility issues are revealed during medical examinations. Common examples include poisoning, nausea, irritable bowel syndrome, gastroesophageal reflux disease (GERD), gas, diarrhea, constipation, and bloating. Structural GID, on the other hand, involves abnormalities in both the appearance and motility of the gastrointestinal tract. Conditions such as strictures, stenosis, inflammatory bowel disease, hemorrhoids, diverticular disease, colorectal polyps, and colorectal cancers fall under this category. GID can be prevented through good health behaviors including regular colonoscopy screenings and good bowel habits [1,6].

1.2. Explainable Artificial Intelligence

Currently, the idea of artificial intelligence is receiving global attention. Artificial intelligence can be defined as “the capability of a machine to imitate intelligent human behavior” (the Merriam–Webster dictionary). It can be denoted that machine learning is a branch of artificial intelligence used “to extract knowledge from large amounts of data” [7]. Popular machine learning methods are support vector machine, random forest, naïve Bayesian predictor, decision tree, and artificial/deep neural networks (see [7] for a description of these methods). The validity of traditional study might be quite limited regarding the early diagnosis of disease, given that it uses logistic regression with an unrealistic assumption of ceteris paribus, i.e., “all the other variables staying constant”. In this context, the emerging literature employs artificial intelligence for the early diagnosis of disease, e.g., breast cancer [8], surgery [9], autism [10], cardiac arrest [11], weight loss [12], blood management [13], depression [14], brain disease [15], preterm birth [16], safe balance [17], neurodevelopmental delay [18], and gait recovery [19]. It is free from unrealistic assumptions of ceteris paribus. It presents the importance values and rankings of independent variables for the early diagnosis of the dependent variable.

Furthermore, the idea of explainable artificial intelligence is gaining great popularity at this point. Explainable artificial intelligence can be denoted as “artificial intelligence to identify major independent variables for the classification or regression of the dependent variable”, and it can be divided into four methods including random forest impurity importance, random forest permutation importance [20,21], machine learning accuracy importance, and Shapley additive explanations (SHAP) [22,23,24,25,26,27,28,29,30,31,32]. Random forest impurity importance is the node impurity decrease from the construction of a branch on a certain independent variable. It is an average over all trees in a random forest with the range of 0 and 1. Random forest permutation importance is the overall accuracy decrease from the permutation of data on the independent variable. It is an average over all trees in the random forest with a value of 0 to 1 [20,21]. Machine learning accuracy importance (an extended version of random forest permutation importance) is the accuracy decrease from the exclusion of data on the independent variable. The SHAP value of an independent variable for a participant is the difference between what machine learning predicts for the probability of the dependent variable (e.g., GID) with and without the independent variable [22,23,24,25,26,27,28,29,30,31,32].

1.3. Genetic Artificial Intelligence

Genetic artificial intelligence (AI) refers to “the application of artificial intelligence for bioinformatics data” [33,34,35]. Bioinformatics combines biology and informatics for the collection, analysis, and interpretation of genetic data. Bioinformatics data include (1) protein sequences consisting of 22 amino acids A, R, N, …, Y, V [proteomics]; (2) deoxyribonucleic acid (DNA) sequences consisting of four nucleotides A, C, G, and T [genomics]; (3) ribonucleic acid (RNA) sequences consisting of four nucleotides A, C, G, and U [genomics]. Here, a nucleotide includes a nitrogenous basis, a pentose sugar, and a phosphate group. In the central dogma of current bioinformatics, DNA sequences are transcribed into RNA sequences, which, finally, are transcribed into protein sequences. In other words, an amino acid of a protein can be expressed as RNA nucleotide triplets (codons) or their DNA counterparts, e.g., Ala (A) as GCU or GCC in

Table 1. RNA and DNA codons.

Amino Acid	RNA Codons	DNA Codons
Ala A	GCU, GCC, GCA, GCG	GCT, GCC, GCA, GCG
Arg R	CGU, CGC, CGA, CGG; AGA, AGG	CGT, CGC, CGA, CGG; AGA, AGG
Asn N	AAU, AAC	AAT, AAC
Asp D	GAU, GAC	GAT, GAC
Asn/Asp B	AAU, AAC; GAU, GAC	AAT, AAC; GAT, GAC
Cys C	UGU, UGC	TGT, TGC
Gln Q	CAA, CAG	CAA, CAG
Glu E	GAA, GAG	GAA, GAG
Gln/Glu Z	CAA, CAG; GAA, GAG	CAA, CAG; GAA, GAG
Gly G	GGU, GGC, GGA, GGG	GGT, GGC, GGA, GGG
His H	CAU, CAC	CAT, CAC
Ile I	AUU, AUC, AUA	ATT, ATC, ATA
Leu L	CUU, CUC, CUA, CUG; UUA, UUG	CTT, CTC, CTA, CTG; TTA, TTG
Lys K	AAA, AAG	AAA, AAG
Met M	AUG	ATG
Phe F	UUU, UUC	TTT, TTC
Pro P	CCU, CCC, CCA, CCG	CCT, CCC, CCA, CCG
Ser S	UCU, UCC, UCA, UCG; AGU, AGC	TCT, TCC, TCA, TCG; AGT, AGC
Thr T	ACU, ACC, ACA, ACG	ACT, ACC, ACA, ACG
Trp W	UGG	TGG
Tyr Y	UAU, UAC	TAT, TAC
Val V	GUU, GUC, GUA, GUG	GTT, GTC, GTA, GTG
Start	AUG, CUG, UUG	ATG, TTG, GTG, CTG
Stop	UAA, UGA, UAG	TAA, TGA, TAG

[36]. For example, insulin, which is a protein chain of 110 amino acids (MALWMR … LENYCN), can be expressed as its DNA counterpart of 110 nucleotide triplets (ATG GCC … TGC AAC). Here, amino acids M, A, C, and N correspond to nucleotide triplets ATG, GCC, TGC, and AAC, respectively. Popular bioinformatics databases are the GenBank [37], GeneCards [38], and UniProt [39].

The human genome is often called “an ocean of information” with 3.2 billion bases. However, only 23,000 genes have been discovered. This accounts for just 3% of the human genome and the rest is still unknown to us. Replication is the quite effective and largely seamless process of copying the DNA sequence in a cell. But a certain error can happen in this process and the cell cannot restore itself completely on some occasions [40]. A single nucleotide polymorphism (SNP) is a one-letter place (A, T, C, or G) where an individual’s DNA sequence varies from other individuals’ DNA sequences [41]. For example, individual 1’s DNA sequence is different, as C, from other individuals’ DNA sequences, as G, in Figure 1 [40]. SNPs can serve as biomarkers, which distinguish those with disease from those without disease. A variation is considered to be an SNP when it occurs in 1% or more of a population. But this “1%” requirement is not universal [41]. An SNP happens after every 1000 nucleotides hence 4 million SNPs occur in the human genome. As addressed above, SNPs happen during the process of replication, and a thousand to a million SNPs occur every single day given that the process of replication never stops in the human genome [40,41]. SNPedia [42] is a common SNP database.

The application of predictive and explainable artificial intelligence for bioinformatics data such as SNP information is attracting rising attention in the diagnosis of various diseases, e.g., asthma [43], diabetes [44], hypertension [45], malaria [46], Parkinson’s diseases [47], schizophrenia [48], and viruses [49,50]. For example, a recent study used the random forest as a predictive and explainable artificial intelligence technique for the classification of asthma [43]. Data came from 128 participants with 176,288 SNPs enrolled in an open data source. The baseline approach (comparison) was the nearest neighbor while the innovation approach (intervention) was the random forest. The intervention (random forest) performed better than the comparison (nearest neighbor) in terms of the accuracy: 62% vs. 49%. The top five SNPs for the prediction of asthma were rs7541950, rs7541956, rs7542025, rs7542028, and rs7542082. This study shows the usefulness of genetic artificial intelligence for the prediction of genetic predisposition to multifactorial diseases like asthma. However, few reviews are available on the recent progress of genetic artificial intelligence for the early diagnosis of GID. In this context, the purpose of this study is to conduct a systematic review for the recent progress of genetic artificial intelligence in GID.

2. Methods

2.1. Data and Search Terms

Figure 2 shows the flow diagram of this study. Data came from ten original studies from PubMed with the following search terms in titles or abstracts: (“snp*” or “single nucleotide polymorphism”) and (“gastro*” or “intestin*” or “diet*” or “digest*” or “stomach*”) and (“machine learning” or “neural network” or “random forest” or “deep learning” or “language model”). The search was conducted during 16 March 2025–15 April 2025.

2.2. Inclusion and Exclusion Criteria

The ten original studies were eligible according to the following criteria: participants with the dependent variable of GID or associated disease; interventions/comparisons of artificial intelligence; outcomes of accuracy, the area under the curve (AUC), and/or variable importance; a publication year of 2010 or later; and the publication language of English. Opinions and reviews were excluded.

2.3. Summary Measures

The following summary measures were adopted: (1) sample size (participants/cases), baseline vs. innovation artificial intelligence methods (comparisons vs. interventions), dependent variable (participants), task type; (2) baseline vs. innovation performance outcomes; (3) major demographic, health-related, and SNP predictors. Here, accuracy denotes the proportion of correct predictions over all observations. The area under the curve (AUC) represents the area under the plot of the true positive rate (sensitivity) against the false positive rate (1-specificity) at various threshold settings. The AUC is a major performance criterion in this study, given that it accommodates sensitivity and specificity.

3. Results

3.1. Summary

A summary of the review for the ten original studies [51,52,53,54,55,56,57,58,59,60] is presented in

Table 2. Summary—sample size, method, and dependent variable.

Study	Sample Size	Method—Baseline	Method—Innovation	Dependent Variable	Type
51	335	RF GERD Included	RF Smoking Included	Esophageal Adenocarcinoma	Classification
52	463	Boosting SNP Excluded	Boosting SNP Included	Crohn’s Disease EIR	Classification
53	3	LR	RF	Gut Microbiome SNP SN	Regression
54	8421	LASSO	LD	Crohn’s Disease	Classification
55	26	RF Microbiome Baseline	RF Microbiome SNP	Colorectal Cancer	Classification
56	1664	ANN	LASSO	GI Nematode Resistance	Classification
57	757,042	ANN	RF	Inflammatory Bowel Disease	Classification
58	570	Wilcoxon Rank Sum	RF	Calcium Metabolism	Classification
59	199,732		Cox	Steatotic Liver Disease	Classification
60	439	RF 1-Year Survival	RF 2-Year Survival	ESCC	Classification

Note. Method: ANN—Artificial Neural Network; LASSO—Least Absolute Shrinkage and Selection Operator; LD—Linkage Disequilibrium; LR—Linear/Logistic Regression; RF—Random Forest; Dependent Variable: ESCC—Esophageal Squamous Cell Carcinoma; EIR—Early Intestinal Resection; GI—Gastrointestinal; SN—Saturation Number.

,

Table 3. Summary—model performance.

Study	Performance—Baseline		Performance—Comparison
	Accuracy	Area Under the Curve	Accuracy	Area Under the Curve
51		70		80
52		81		84
53		80		99
54		52		63
55	87		92
56	65		79
57		91		98
58	25		100
59				99
60		65		80
Min	25	52	79	63
Max	87	91	100	99
	R-Square
	100 × (1 − p value)

and

Table 4. Summary—major predictor.

Study	Predictor Demographic	Predictor Health	Predictor SNP
51		GERD Smoking BMI	rs2295778	rs13337626	rs2296188	rs2114039
			rs11941492	rs17708574	rs7324547	rs17619601
			rs17625898
52	Age	Disease Behavior	rs28785174	rs60532570	rs13056955	rs7660164
53		SNP Density/Number
54		MARCKS Protein	rs4945943
55		Microbiome Intestinal	IB175794	BA459738	EM8439
56			SNPs 41676
57		Microbiome Intestinal	SNPs 13	EXOC3: 6	SLC25A26: 1	YIF1B: 6
58		Microbiome Intestinal	rs316115020	rs316420452
59			rs738409_G	rs2642438_A	rs58542926_T	rs72613567_TA
60		Immune Cell Types	rs148710154	rs75146099

Note. Predictor Health: BMI—Body Mass Index; GERD—Gastroesophageal Reflux Disease; Predictor SNP (Single Nucleotide Polymorphism): BA—Blautia_A sp900066145; EM—Eubacterium_M sp; IB—Intestinibacter Bartlettii; SNP—Single Nucleotide Polymorphism.

. The “Study” column in the tables denotes the reference numbers of the ten original studies. The tables include (1) sample size (participants/cases), baseline vs. innovation artificial intelligence methods, dependent variable and task type (Table 2); (2) baseline vs. innovation performance outcomes (Table 3); (3) major demographic, health-related, and SNP predictors (Table 4). The performance outcomes reported were within 79–100 for accuracy (%) and 63–98 for the AUC (%). The random forest registered the best performance (AUC 98%) for the classifications of inflammatory bowel disease with 13 single nucleotide polymorphisms (SNPs). Likewise, the random forest delivered the best performance (R-square 99%) for the regression of the gut microbiome SNP saturation number. The following SNPs were discovered to be major predictors of GID or associated disease: rs2295778, rs13337626, rs2296188, rs2114039 (esophageal adenocarcinoma); rs28785174, rs60532570, rs13056955, rs7660164 (Crohn’s disease early intestinal resection); rs4945943 (Crohn’s disease); rs316115020, rs316420452 (calcium metabolism); rs738409_G, rs2642438_A, rs58542926_T, rs72613567_TA (steatotic liver disease); and rs148710154, rs75146099 (esophageal squamous cell carcinoma). The following demographic and health-related variables were found to be important predictors of GID or associated disease besides SNPs: age, body mass index, disease behavior, immune cell type, intestinal microbiome, MARCKS protein, smoking, and SNP density/number. No deep learning study was found even though deep learning was used as a search term together with machine learning. However, artificial intelligence is a data-driven approach, and more research is needed for more general conclusions.

3.2. Genetic Artificial Intelligence for Inflammatory Bowel Disease

This section describes three original studies of genetic artificial intelligence for inflammatory bowel disease [52,54,57]. A recent study used boosting as a predictive and explainable artificial intelligence technique for the classification of Crohn’s disease early intestinal dissection [52]. The data came from 463 participants enrolled in 15 general hospitals in Korea. The baseline approach (comparison) was baseline boosting (excluding SNPs) while the innovation approach (intervention) was genetic boosting (including SNPs). The intervention (genetic boosting) performed better than the comparison (baseline boosting) in terms of the AUC: 84% vs. 81%. Age and disease behavior were major features of genetic boosting together with four SNPs, i.e., rs28785174, rs60532570, rs13056955, and rs7660164. Another recent study adopted Linkage Disequilibrium (LD) as the predictive and explainable machine learning technique for the classification of Crohn’s disease [54]. The data source was 8421 cases enrolled in the European Genome-phenome Archive. The Least Absolute Shrinkage and Selection Operator (LASSO) served as the comparison whereas the LD served as the intervention. The intervention (LD) surpassed the comparison (LASSO) in terms of the AUC: 63% vs. 52%. One SNP (rs4945943) and MARCKS protein were important features of LD for the prediction of Crohn’s disease. Likewise, a recent study employed the random forest as a predictive and explainable machine learning technique for the classification of inflammatory bowel disease [57]. The data consisted of 757,042 cases enrolled in genome-wide association studies for European populations. The artificial neural network and the random forest were considered as the comparison and the intervention, respectively. The AUC of the intervention (random forest) was better than that of the comparison (artificial neural network), i.e., 98% vs. 91%. Thirteen SNPs (e.g., EXOC3: 6, SLC25A26: 1, YIF1B: 6) and intestinal microbiomes were significant features of the random forest for the prediction of inflammatory bowel disease.

3.3. Genetic Artificial Intelligence for Gastrointestinal Cancer

This section focuses on three original studies based on genetic artificial intelligence for gastrointestinal cancer [51,55,60]. A recent study used the random forest as a predictive and explainable artificial intelligence technique for the classification of esophageal adenocarcinoma [51]. The data came from 335 participants enrolled in a general hospital in the United States. The baseline approach (comparison) was the random forest including gastroesophageal reflux disease (RF-GERD) while the innovation approach (intervention) was the random forest including smoking (RF-Smoking). The intervention (RF-Smoking) went beyond the comparison (RF-GERD) in terms of the AUC, i.e., 80% vs. 70%. Besides smoking, nine SNPs were major predictors of RF-Smoking for esophageal adenocarcinoma: rs2295778, rs13337626, rs2296188, rs2114039, rs11941492, rs17708574, rs7324547, rs17619601, and rs17625898. Another recent study adopted the random forest as the predictive and explainable machine learning technique for the classification of colorectal cancer [55]. The data source was 26 cases enrolled in a general hospital in China. The random forest including microbiome (RF Microbiome Baseline) served as the comparison whereas the random forest including microbiome SNPs (RF Microbiome SNP) served as the intervention. The intervention (RF Microbiome SNP) surpassed the comparison (RF Microbiome Baseline) in terms of accuracy: 92% vs. 87%. Here, IB175794, BA459738, and EM8439 were found to be important intestinal microbiome SNPs. In a similar context, a recent study employed the random forest as the predictive and explainable machine learning technique for the classification of esophageal squamous cell carcinoma survival [60]. The data consisted of 439 participants enrolled in the Gene Expression Omnibus database. The random forest for 1-year survival and the random forest for 2-year survival were considered as the comparison and the intervention, respectively. The C-index of the intervention (random forest for 2-year survival) was beyond that of the comparison (random forest for 1-year survival), i.e., 80% vs. 65%. Two SNPs (rs148710154, rs75146099) and immune cell type were significant predictors of the random forest for esophageal squamous cell carcinoma survival.

3.4. Genetic Artificial Intelligence for Other Gastrointestinal Diseases

This section highlights three original studies based on genetic artificial intelligence for other gastrointestinal diseases [53,56,58]. A recent study used the random forest as the predictive and explainable artificial intelligence technique for the regression of gut microbiome SNP saturation number [53]. The data came from three participants enrolled in the European Nucleotide Archive. The baseline approach (comparison) was linear regression while the innovation approach (intervention) was the random forest. The intervention (random forest) was superior to the comparison (linear regression) in terms of the R-square: 99% vs. 80%. SNP density/number was a major variable of the random forest for the regression of gut microbiome SNP saturation number. Another recent study adopted the LASSO as the predictive and explainable machine learning technique for the classification of GI nematodes [56]. The data source was 1664 cases enrolled in Brazil. The artificial neural network served as the comparison whereas the LASSO served as the intervention. The intervention (LASSO) surpassed the comparison (artificial neural network) in terms of accuracy: 79% vs. 65%. A total of 41,676 SNPs were important variables of LASSO for the prediction of GI nematodes. In a similar vein, a recent study employed the random forest as the predictive and explainable machine learning technique for the classification of calcium metabolism [58]. The data consisted of 570 cases enrolled in China. The Wilcoxon test and random forest were considered as the comparison and the intervention, respectively. The accuracy of the intervention (random forest) was higher than that of the comparison (Wilcoxon’s test), i.e., 100% vs. 25%. Two SNPs (rs148710154, rs75146099) and intestinal microbiomes were significant variables of random forest for the prediction of calcium metabolism.

4. Discussion

4.1. Contributions of This Study

As addressed above, the application of predictive and explainable artificial intelligence to bioinformatics data such as SNP information is registering a rapid expansion in the diagnosis of various diseases [43,44,45,46,47,48,49,50]. However, little review has been conducted on the recent progress of genetic artificial intelligence for the early diagnosis of GID. In this vein, this study conducted a systematic review for the recent progress of genetic artificial intelligence in GID. One potential contribution of this review would be for pharmacology, which is the systematic study of drugs and their effects on living organisms. In other words, it focuses on the interactions between chemical and biological systems. It covers the chemical compositions, biological mechanisms, therapeutic uses, and potential toxicities of drugs. Here, SNPs can become drug targets or affect drug response for various diseases such as diabetes, obesity, and psychiatric disorders [61]. As addressed above, SNPs can serve as biomarkers, which distinguish those with disease from those without disease. Also, SNPs within transcription factor binding sites can influence the transcription rates of target genes thereby inducing a certain disease or its phenotypes [62]. This would be true for GID as well. SNPs can become drug targets or affect drug response based on GID proteomics and genomics.

4.2. Limitations of Existing Literature

Previous studies on the early diagnosis of GID based on explainable artificial intelligence had some limitations. Firstly, the four approaches of explainable artificial intelligence at this point—random forest impurity importance, random forest permutation importance, machine learning accuracy importance, and SHAP—can yield different results in certain circumstances. Random forest impurity importance can vary depending on how variables are categorized, while random forest permutation importance is relatively unaffected by this potential variation [21]. It should be noted, however, that random forest is unique in considering sequential information and this unique strength becomes more apparent with its impurity importance than with its permutation importance. In this vein, it is highly recommended to compare the four approaches of explainable artificial intelligence in a comprehensive manner. Secondly, it was beyond the scope of this review to consider other types of explainable artificial intelligence such as local interpretable model-agnostic explanations (LIME) [63]. Thirdly, the AUC of a study (0.63) [54] was below the optimal level for a diagnostic test.

4.3. Suggestions for Future Research

Some suggestions for this line of research are presented here. Firstly, uniting various modes of explainable artificial intelligence for various modes of GID data would give deeper clinical insights. For instance, one recent study [64] developed the random forest prediction system of oral cancer survival based on pathological (image), genetic, and clinical (numeric) data. The multi-modal random forest was far beyond other approaches in terms of model performance (c-index), i.e., 83% vs. multi-modal boosting (75%), multi-modal Cox (74%), clinical random forest (70%), genetic random forest (64%), and pathological random forest (64%). Here, it can be noted that clinical, genetic, and pathological predictors were equally important in unimodal models. In addition, protein binding was highly enriched based on gene enrichment analysis, whereas plasma membranes, secreted proteins, and extracellular regions were highly represented according to cellular component analysis. However, little literature is available, and more examination is needed regarding the combination of various modes of explainable artificial intelligence for various modes of GID data.

Secondly, more rigorous qualitative evaluation approaches are needed regarding systematic reviews of genetic artificial intelligence in GID. The Enhancing the Quality and Transparency of Health Research Network suggests the inclusion of a research question; eligibility and exclusion criteria; flow diagram; and experimental characteristics including sample size (participants), baseline vs. innovation methods (comparisons vs. interventions), dependent variable (participants), task type, baseline vs. innovation performance outcomes, and participant characteristics [65,66]. This study adopted the following suggestions: research question (Section 1); eligibility and exclusion criteria (Section 2); flow diagram (Figure 2); and experimental characteristics such as sample size, baseline vs. innovation artificial intelligence methods, dependent variable, task type (Table 2), baseline vs. innovation performance outcomes (Table 3), and major demographic, health-related, and SNP predictors (Table 4). But, more rigorous qualitative evaluation approaches can be introduced and this new guideline is likely to strengthen the validity of reviews for genetic artificial intelligence in GID.

Thirdly, little literature has been available and more investigation is needed on genetic artificial intelligence for reinforcement learning. Reinforcement learning is a division of artificial intelligence in which the environment gives a series of rewards, an agent responds with a series of actions to maximize the cumulative reward, and the environment makes a transition to the next period with given probabilities [67]. Reinforcement learning artificial intelligence begins like a human player with limited information in the limited periods available but eventually surpasses the best human player ever with the sheer power of big data (“temporal difference learning” in a professional language) [67]. Reinforcement learning has enjoyed great success in finance [68] and healthcare [69,70,71]. However, little research has been conducted and more examination is needed on explainable reinforcement learning. Reportedly, there have been a few studies in this direction and these studies have centered on simplified models with easy interpretation but insufficient performance and little consideration of the psychological and social factors behind optimization processes [72].

4.4. Conclusions

In summary, this study reviewed the recent advances in genetic artificial intelligence for the early diagnosis of GID. The performance results reported were within 79–100 for accuracy (%) and 63–98 for the AUC (%). Random forest presented the best performance (AUC 98%) for the classifications of inflammatory bowel disease with 13 single nucleotide polymorphisms (SNPs). Similarly, random forest showed the best performance (R-square 99%) for the regression of the gut microbiome SNP saturation number. The following SNPs were found to be major variables of GID or associated disease: rs2295778, rs13337626, rs2296188, rs2114039 (esophageal adenocarcinoma); rs28785174, rs60532570, rs13056955, rs7660164 (Crohn’s disease early intestinal resection); rs4945943 (Crohn’s disease); rs316115020, rs316420452 (calcium metabolism); rs738409_G, rs2642438_A, rs58542926_T, rs72613567_TA (steatotic liver disease); and rs148710154, rs75146099 (esophageal squamous cell carcinoma). The following demographic and health-related predictors were discovered to be important variables of GID or associated disease besides SNPs: age, body mass index, disease behavior, immune cell type, intestinal microbiome, MARCKS protein, smoking, and SNP density/number. In conclusion, genetic artificial intelligence is effective and non-invasive as a decision support system for GID.